Know what each customer costs you.
Monthly averages can make an AI feature look profitable while a small group of users consumes most of its budget. Measure usage by customer so your pricing has something solid underneath it.
Add up the variable costs
Include model calls, embeddings, storage, search, tool execution, and retries. Attribute shared infrastructure sensibly. Separate one-time onboarding work from recurring activity so your first-month figures do not distort the ongoing margin.
Look beyond the average
Compare typical usage with the busiest customers. A plan that works at the median may fail at the upper end. Test what happens when users send larger files, request longer responses, or use a feature more frequently.
Give the product a budget
Set usage allowances and clear limits where needed. Review cost per completed task and contribution margin as model rates change. Credits can soften a bill, but temporary discounts should not be the only reason your pricing works.